#include "llvm/Transforms/Scalar.h"
#include "llvm/IntrinsicInst.h"
#include "llvm/Instructions.h"
+#include "llvm/LLVMContext.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/Dominators.h"
#include "llvm/Analysis/MemoryDependenceAnalysis.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
+#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetData.h"
#include <list>
using namespace llvm;
STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted");
STATISTIC(NumMemSetInfer, "Number of memsets inferred");
+STATISTIC(NumMoveToCpy, "Number of memmoves converted to memcpy");
/// isBytewiseValue - If the specified value can be set by repeating the same
/// byte in memory, return the i8 value that it is represented with. This is
/// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated
/// byte store (e.g. i16 0x1234), return null.
static Value *isBytewiseValue(Value *V) {
+ LLVMContext &Context = V->getContext();
+
// All byte-wide stores are splatable, even of arbitrary variables.
- if (V->getType() == Type::Int8Ty) return V;
+ if (V->getType()->isIntegerTy(8)) return V;
// Constant float and double values can be handled as integer values if the
// corresponding integer value is "byteable". An important case is 0.0.
if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
- if (CFP->getType() == Type::FloatTy)
- V = ConstantExpr::getBitCast(CFP, Type::Int32Ty);
- if (CFP->getType() == Type::DoubleTy)
- V = ConstantExpr::getBitCast(CFP, Type::Int64Ty);
+ if (CFP->getType()->isFloatTy())
+ V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(Context));
+ if (CFP->getType()->isDoubleTy())
+ V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(Context));
// Don't handle long double formats, which have strange constraints.
}
if (Val != Val2)
return 0;
}
- return ConstantInt::get(Val);
+ return ConstantInt::get(Context, Val);
}
}
// Otherwise, we have a sequential type like an array or vector. Multiply
// the index by the ElementSize.
- uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
+ uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
Offset += Size*OpC->getSExtValue();
}
if (Start < I->Start) {
I->Start = Start;
I->StartPtr = SI->getPointerOperand();
+ I->Alignment = SI->getAlignment();
}
// Now we know that Start <= I->End and Start >= I->Start (so the startpoint
// End.
if (End > I->End) {
I->End = End;
- range_iterator NextI = I;;
+ range_iterator NextI = I;
while (++NextI != E && End >= NextI->Start) {
// Merge the range in.
I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end());
//===----------------------------------------------------------------------===//
namespace {
-
- class VISIBILITY_HIDDEN MemCpyOpt : public FunctionPass {
+ class MemCpyOpt : public FunctionPass {
bool runOnFunction(Function &F);
public:
static char ID; // Pass identification, replacement for typeid
AU.addRequired<DominatorTree>();
AU.addRequired<MemoryDependenceAnalysis>();
AU.addRequired<AliasAnalysis>();
- AU.addRequired<TargetData>();
AU.addPreserved<AliasAnalysis>();
AU.addPreserved<MemoryDependenceAnalysis>();
- AU.addPreserved<TargetData>();
}
// Helper fuctions
- bool processStore(StoreInst *SI, BasicBlock::iterator& BBI);
- bool processMemCpy(MemCpyInst* M);
- bool performCallSlotOptzn(MemCpyInst* cpy, CallInst* C);
+ bool processStore(StoreInst *SI, BasicBlock::iterator &BBI);
+ bool processMemCpy(MemCpyInst *M);
+ bool processMemMove(MemMoveInst *M);
+ bool performCallSlotOptzn(MemCpyInst *cpy, CallInst *C);
bool iterateOnFunction(Function &F);
};
// createMemCpyOptPass - The public interface to this file...
FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOpt(); }
-static RegisterPass<MemCpyOpt> X("memcpyopt",
- "MemCpy Optimization");
+INITIALIZE_PASS(MemCpyOpt, "memcpyopt", "MemCpy Optimization", false, false);
/// some other patterns to fold away. In particular, this looks for stores to
/// neighboring locations of memory. If it sees enough consequtive ones
/// (currently 4) it attempts to merge them together into a memcpy/memset.
-bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator& BBI) {
+bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
if (SI->isVolatile()) return false;
+ LLVMContext &Context = SI->getContext();
+
// There are two cases that are interesting for this code to handle: memcpy
// and memset. Right now we only handle memset.
if (!ByteVal)
return false;
- TargetData &TD = getAnalysis<TargetData>();
+ TargetData *TD = getAnalysisIfAvailable<TargetData>();
+ if (!TD) return false;
AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
+ Module *M = SI->getParent()->getParent()->getParent();
// Okay, so we now have a single store that can be splatable. Scan to find
// all subsequent stores of the same value to offset from the same pointer.
// Join these together into ranges, so we can decide whether contiguous blocks
// are stored.
- MemsetRanges Ranges(TD);
+ MemsetRanges Ranges(*TD);
Value *StartPtr = SI->getPointerOperand();
// Check to see if this store is to a constant offset from the start ptr.
int64_t Offset;
- if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), Offset, TD))
+ if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), Offset, *TD))
break;
Ranges.addStore(Offset, NextStore);
// store as well. We try to avoid this unless there is at least something
// interesting as a small compile-time optimization.
Ranges.addStore(0, SI);
-
- Function *MemSetF = 0;
// Now that we have full information about ranges, loop over the ranges and
// emit memset's for anything big enough to be worthwhile.
if (Range.TheStores.size() == 1) continue;
// If it is profitable to lower this range to memset, do so now.
- if (!Range.isProfitableToUseMemset(TD))
+ if (!Range.isProfitableToUseMemset(*TD))
continue;
// Otherwise, we do want to transform this! Create a new memset. We put
// memset block. This ensure that the memset is dominated by any addressing
// instruction needed by the start of the block.
BasicBlock::iterator InsertPt = BI;
-
- if (MemSetF == 0) {
- const Type *Tys[] = {Type::Int64Ty};
- MemSetF = Intrinsic::getDeclaration(SI->getParent()->getParent()
- ->getParent(), Intrinsic::memset,
- Tys, 1);
- }
-
+
// Get the starting pointer of the block.
StartPtr = Range.StartPtr;
-
+
+ // Determine alignment
+ unsigned Alignment = Range.Alignment;
+ if (Alignment == 0) {
+ const Type *EltType =
+ cast<PointerType>(StartPtr->getType())->getElementType();
+ Alignment = TD->getABITypeAlignment(EltType);
+ }
+
// Cast the start ptr to be i8* as memset requires.
- const Type *i8Ptr = PointerType::getUnqual(Type::Int8Ty);
- if (StartPtr->getType() != i8Ptr)
- StartPtr = new BitCastInst(StartPtr, i8Ptr, StartPtr->getNameStart(),
+ const PointerType* StartPTy = cast<PointerType>(StartPtr->getType());
+ const PointerType *i8Ptr = Type::getInt8PtrTy(Context,
+ StartPTy->getAddressSpace());
+ if (StartPTy!= i8Ptr)
+ StartPtr = new BitCastInst(StartPtr, i8Ptr, StartPtr->getName(),
InsertPt);
-
+
Value *Ops[] = {
StartPtr, ByteVal, // Start, value
- ConstantInt::get(Type::Int64Ty, Range.End-Range.Start), // size
- ConstantInt::get(Type::Int32Ty, Range.Alignment) // align
+ // size
+ ConstantInt::get(Type::getInt64Ty(Context), Range.End-Range.Start),
+ // align
+ ConstantInt::get(Type::getInt32Ty(Context), Alignment),
+ // volatile
+ ConstantInt::get(Type::getInt1Ty(Context), 0),
};
- Value *C = CallInst::Create(MemSetF, Ops, Ops+4, "", InsertPt);
- DEBUG(cerr << "Replace stores:\n";
+ const Type *Tys[] = { Ops[0]->getType(), Ops[2]->getType() };
+
+ Function *MemSetF = Intrinsic::getDeclaration(M, Intrinsic::memset, Tys, 2);
+
+ Value *C = CallInst::Create(MemSetF, Ops, Ops+5, "", InsertPt);
+ DEBUG(dbgs() << "Replace stores:\n";
for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i)
- cerr << *Range.TheStores[i];
- cerr << "With: " << *C); C=C;
+ dbgs() << *Range.TheStores[i];
+ dbgs() << "With: " << *C); C=C;
// Don't invalidate the iterator
BBI = BI;
// Zap all the stores.
- for (SmallVector<StoreInst*, 16>::const_iterator SI = Range.TheStores.begin(),
+ for (SmallVector<StoreInst*, 16>::const_iterator
+ SI = Range.TheStores.begin(),
SE = Range.TheStores.end(); SI != SE; ++SI)
(*SI)->eraseFromParent();
++NumMemSetInfer;
// Deliberately get the source and destination with bitcasts stripped away,
// because we'll need to do type comparisons based on the underlying type.
- Value* cpyDest = cpy->getDest();
- Value* cpySrc = cpy->getSource();
+ Value *cpyDest = cpy->getDest();
+ Value *cpySrc = cpy->getSource();
CallSite CS = CallSite::get(C);
// We need to be able to reason about the size of the memcpy, so we require
// that it be a constant.
- ConstantInt* cpyLength = dyn_cast<ConstantInt>(cpy->getLength());
+ ConstantInt *cpyLength = dyn_cast<ConstantInt>(cpy->getLength());
if (!cpyLength)
return false;
// Require that src be an alloca. This simplifies the reasoning considerably.
- AllocaInst
* srcAlloca = dyn_cast<AllocaInst>(cpySrc);
+ AllocaInst
*srcAlloca = dyn_cast<AllocaInst>(cpySrc);
if (!srcAlloca)
return false;
// Check that all of src is copied to dest.
- TargetData& TD = getAnalysis<TargetData>();
+ TargetData *TD = getAnalysisIfAvailable<TargetData>();
+ if (!TD) return false;
- ConstantInt* srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
+ ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
if (!srcArraySize)
return false;
- uint64_t srcSize = TD.getABITypeSize(srcAlloca->getAllocatedType()) *
+ uint64_t srcSize = TD->getTypeAllocSize(srcAlloca->getAllocatedType()) *
srcArraySize->getZExtValue();
if (cpyLength->getZExtValue() < srcSize)
// Check that accessing the first srcSize bytes of dest will not cause a
// trap. Otherwise the transform is invalid since it might cause a trap
// to occur earlier than it otherwise would.
- if (AllocaInst
* A = dyn_cast<AllocaInst>(cpyDest)) {
+ if (AllocaInst
*A = dyn_cast<AllocaInst>(cpyDest)) {
// The destination is an alloca. Check it is larger than srcSize.
- ConstantInt* destArraySize = dyn_cast<ConstantInt>(A->getArraySize());
+ ConstantInt *destArraySize = dyn_cast<ConstantInt>(A->getArraySize());
if (!destArraySize)
return false;
- uint64_t destSize = TD.getABITypeSize(A->getAllocatedType()) *
+ uint64_t destSize = TD->getTypeAllocSize(A->getAllocatedType()) *
destArraySize->getZExtValue();
if (destSize < srcSize)
return false;
- } else if (Argument
* A = dyn_cast<Argument>(cpyDest)) {
+ } else if (Argument
*A = dyn_cast<Argument>(cpyDest)) {
// If the destination is an sret parameter then only accesses that are
// outside of the returned struct type can trap.
if (!A->hasStructRetAttr())
return false;
- const Type
* StructTy = cast<PointerType>(A->getType())->getElementType();
- uint64_t destSize = TD.getABITypeSize(StructTy);
+ const Type
*StructTy = cast<PointerType>(A->getType())->getElementType();
+ uint64_t destSize = TD->getTypeAllocSize(StructTy);
if (destSize < srcSize)
return false;
SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(),
srcAlloca->use_end());
while (!srcUseList.empty()) {
- User* UI = srcUseList.back();
- srcUseList.pop_back();
+ User *UI = srcUseList.pop_back_val();
if (isa<BitCastInst>(UI)) {
for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
I != E; ++I)
srcUseList.push_back(*I);
- } else if (GetElementPtrInst* G = dyn_cast<GetElementPtrInst>(UI)) {
+ } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(UI)) {
if (G->hasAllZeroIndices())
for (User::use_iterator I = UI->use_begin(), E = UI->use_end();
I != E; ++I)
// Since we're changing the parameter to the callsite, we need to make sure
// that what would be the new parameter dominates the callsite.
- DominatorTree& DT = getAnalysis<DominatorTree>();
- if (Instruction* cpyDestInst = dyn_cast<Instruction>(cpyDest))
+ DominatorTree &DT = getAnalysis<DominatorTree>();
+ if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest))
if (!DT.dominates(cpyDestInst, C))
return false;
// unexpected manner, for example via a global, which we deduce from
// the use analysis, we also need to know that it does not sneakily
// access dest. We rely on AA to figure this out for us.
- AliasAnalysis
& AA = getAnalysis<AliasAnalysis>();
+ AliasAnalysis
&AA = getAnalysis<AliasAnalysis>();
if (AA.getModRefInfo(C, cpy->getRawDest(), srcSize) !=
AliasAnalysis::NoModRef)
return false;
cpyDest = CastInst::CreatePointerCast(cpyDest, cpySrc->getType(),
cpyDest->getName(), C);
changedArgument = true;
- if (CS.getArgument(i)->getType() != cpyDest->getType())
- CS.setArgument(i, CastInst::CreatePointerCast(cpyDest,
- CS.getArgument(i)->getType(), cpyDest->getName(), C));
- else
+ if (CS.getArgument(i)->getType() == cpyDest->getType())
CS.setArgument(i, cpyDest);
+ else
+ CS.setArgument(i, CastInst::CreatePointerCast(cpyDest,
+ CS.getArgument(i)->getType(), cpyDest->getName(), C));
}
if (!changedArgument)
// Drop any cached information about the call, because we may have changed
// its dependence information by changing its parameter.
- MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
- MD.dropInstruction(C);
+ MemoryDependenceAnalysis &MD = getAnalysis<MemoryDependenceAnalysis>();
+ MD.removeInstruction(C);
// Remove the memcpy
MD.removeInstruction(cpy);
cpy->eraseFromParent();
- NumMemCpyInstr++;
+ ++NumMemCpyInstr;
return true;
}
/// copies X to Y, and memcpy B which copies Y to Z, then we can rewrite B to be
/// a memcpy from X to Z (or potentially a memmove, depending on circumstances).
/// This allows later passes to remove the first memcpy altogether.
-bool MemCpyOpt::processMemCpy(MemCpyInst* M) {
- MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>();
+bool MemCpyOpt::processMemCpy(MemCpyInst *M) {
+ MemoryDependenceAnalysis &MD = getAnalysis<MemoryDependenceAnalysis>();
// The are two possible optimizations we can do for memcpy:
- // a) memcpy-memcpy xform which exposes redundance for DSE
- // b) call-memcpy xform for return slot optimization
+ // a) memcpy-memcpy xform which exposes redundance for DSE.
+ // b) call-memcpy xform for return slot optimization.
MemDepResult dep = MD.getDependency(M);
- if (!dep.isNormal())
+ if (!dep.isClobber())
return false;
- else if (!isa<MemCpyInst>(dep.getInst())) {
- if (CallInst* C = dyn_cast<CallInst>(dep.getInst()))
+ if (!isa<MemCpyInst>(dep.getInst())) {
+ if (CallInst *C = dyn_cast<CallInst>(dep.getInst()))
return performCallSlotOptzn(M, C);
- else
- return false;
+ return false;
}
- MemCpyInst* MDep = cast<MemCpyInst>(dep.getInst());
+ MemCpyInst *MDep = cast<MemCpyInst>(dep.getInst());
// We can only transforms memcpy's where the dest of one is the source of the
// other
// Second, the length of the memcpy's must be the same, or the preceeding one
// must be larger than the following one.
- ConstantInt* C1 = dyn_cast<ConstantInt>(MDep->getLength());
- ConstantInt* C2 = dyn_cast<ConstantInt>(M->getLength());
+ ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
+ ConstantInt *C2 = dyn_cast<ConstantInt>(M->getLength());
if (!C1 || !C2)
return false;
// Finally, we have to make sure that the dest of the second does not
// alias the source of the first
- AliasAnalysis
& AA = getAnalysis<AliasAnalysis>();
+ AliasAnalysis
&AA = getAnalysis<AliasAnalysis>();
if (AA.alias(M->getRawDest(), CpySize, MDep->getRawSource(), DepSize) !=
AliasAnalysis::NoAlias)
return false;
return false;
// If all checks passed, then we can transform these memcpy's
- const Type *Tys[1];
- Tys[0] = M->getLength()->getType();
- Function* MemCpyFun = Intrinsic::getDeclaration(
+ const Type *ArgTys[3] = { M->getRawDest()->getType(),
+ MDep->getRawSource()->getType(),
+ M->getLength()->getType() };
+ Function *MemCpyFun = Intrinsic::getDeclaration(
M->getParent()->getParent()->getParent(),
- M->getIntrinsicID(), Tys, 1);
+ M->getIntrinsicID(), ArgTys, 3);
- std::vector<Value*> args;
- args.push_back(M->getRawDest());
- args.push_back(MDep->getRawSource());
- args.push_back(M->getLength());
- args.push_back(M->getAlignment());
+ Value *Args[5] = {
+ M->getRawDest(), MDep->getRawSource(), M->getLength(),
+ M->getAlignmentCst(), M->getVolatileCst()
+ };
- CallInst* C = CallInst::Create(MemCpyFun, args.begin(), args.end(), "", M);
+ CallInst *C = CallInst::Create(MemCpyFun, Args, Args+5, "", M);
// If C and M don't interfere, then this is a valid transformation. If they
// did, this would mean that the two sources overlap, which would be bad.
if (MD.getDependency(C) == dep) {
- MD.dropInstruction(M);
+ MD.removeInstruction(M);
M->eraseFromParent();
-
- NumMemCpyInstr++;
-
+ ++NumMemCpyInstr;
return true;
}
// inserted and act like nothing happened.
MD.removeInstruction(C);
C->eraseFromParent();
-
return false;
}
-// MemCpyOpt::runOnFunction - This is the main transformation entry point for a
-// function.
-//
-bool MemCpyOpt::runOnFunction(Function& F) {
+/// processMemMove - Transforms memmove calls to memcpy calls when the src/dst
+/// are guaranteed not to alias.
+bool MemCpyOpt::processMemMove(MemMoveInst *M) {
+ AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
+
+ // If the memmove is a constant size, use it for the alias query, this allows
+ // us to optimize things like: memmove(P, P+64, 64);
+ uint64_t MemMoveSize = ~0ULL;
+ if (ConstantInt *Len = dyn_cast<ConstantInt>(M->getLength()))
+ MemMoveSize = Len->getZExtValue();
- bool changed = false;
- bool shouldContinue = true;
+ // See if the pointers alias.
+ if (AA.alias(M->getRawDest(), MemMoveSize, M->getRawSource(), MemMoveSize) !=
+ AliasAnalysis::NoAlias)
+ return false;
- while (shouldContinue) {
- shouldContinue = iterateOnFunction(F);
- changed |= shouldContinue;
- }
+ DEBUG(dbgs() << "MemCpyOpt: Optimizing memmove -> memcpy: " << *M << "\n");
- return changed;
-}
+ // If not, then we know we can transform this.
+ Module *Mod = M->getParent()->getParent()->getParent();
+ const Type *ArgTys[3] = { M->getRawDest()->getType(),
+ M->getRawSource()->getType(),
+ M->getLength()->getType() };
+ M->setCalledFunction(Intrinsic::getDeclaration(Mod, Intrinsic::memcpy, ArgTys, 3));
+ // MemDep may have over conservative information about this instruction, just
+ // conservatively flush it from the cache.
+ getAnalysis<MemoryDependenceAnalysis>().removeInstruction(M);
-// MemCpyOpt::iterateOnFunction - Executes one iteration of GVN
+ ++NumMoveToCpy;
+ return true;
+}
+
+
+// MemCpyOpt::iterateOnFunction - Executes one iteration of GVN.
bool MemCpyOpt::iterateOnFunction(Function &F) {
- bool changed_function = false;
+ bool MadeChange = false;
- // Walk all instruction in the function
+ // Walk all instruction in the function.
for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) {
for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
BI != BE;) {
- // Avoid invalidating the iterator
- Instruction* I = BI++;
+ // Avoid invalidating the iterator.
+ Instruction *I = BI++;
if (StoreInst *SI = dyn_cast<StoreInst>(I))
- changed_function |= processStore(SI, BI);
- else if (MemCpyInst* M = dyn_cast<MemCpyInst>(I)) {
- changed_function |= processMemCpy(M);
+ MadeChange |= processStore(SI, BI);
+ else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I))
+ MadeChange |= processMemCpy(M);
+ else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I)) {
+ if (processMemMove(M)) {
+ --BI; // Reprocess the new memcpy.
+ MadeChange = true;
+ }
}
}
}
- return changed_function;
+ return MadeChange;
+}
+
+// MemCpyOpt::runOnFunction - This is the main transformation entry point for a
+// function.
+//
+bool MemCpyOpt::runOnFunction(Function &F) {
+ bool MadeChange = false;
+ while (1) {
+ if (!iterateOnFunction(F))
+ break;
+ MadeChange = true;
+ }
+
+ return MadeChange;
}
+
+
+
projects / oota-llvm.git / blobdiff